Spurious Correlation Robustness

Despite substantial advances, large language models (LLMs) continue to exhibit hallucinations, generating plausible yet incorrect responses. In this paper, we highlight a critical yet previously underexplored class of hallucinations driven by spurious correlations -- superficial but statistically prominent associations between features (e.g., surnames) and attributes (e.g., nationality) present in the training data. We demonstrate that these spurious correlations induce hallucinations that are confidently generated, immune to model scaling, evade current detection methods, and persist even after refusal fine-tuning. Through systematically controlled synthetic experiments and empirical evaluations on state-of-the-art open-source and proprietary LLMs (including GPT-5), we show that existing hallucination detection methods, such as confidence-based filtering and inner-state probing, fundamentally fail in the presence of spurious correlations. Our theoretical analysis further elucidates why these statistical biases intrinsically undermine confidence-based detection techniques. Our findings thus emphasize the urgent need for new approaches explicitly designed to address hallucinations caused by spurious correlations.

Predictors often exploit spurious correlations that fail under distribution shift, undermining out-of-distribution (OOD) robustness. Invariant pairs, such as the same object photographed against different backgrounds, directly indicate which variation should not affect prediction. Yet real pair collections are finite and noisy, so they can miss relevant variation or lead a learner to discard predictive information. To learn from such collections, we propose Geometric Robustness Invariant Training (GRIT), which estimates an unstable subspace from the covariance of pair differences in fixed pretrained features and projects it out before fitting a predictor. Because pair processing is separate from predictor fitting, the same projection can accompany ERM or other robustness objectives. Unlike quadratic prediction consistency, which penalizes variation along every direction with nonzero pair covariance, including variation introduced by pair noise, GRIT enforces exact invariance only to the estimated subspace. We bound GRIT's target risk in terms of pair count, noise relative to the weakest clean signal, and selected rank, assuming that the label distribution given a linear projection of the features is invariant across environments. Empirically, GRIT improves the mean OOD accuracy of each of eight learning objectives on ColoredMNIST-INV and Waterbirds-INV and has higher mean accuracy than a prediction-consistency penalty using the same pairs in 15 of 16 settings. On UrbanCars, combining pairs that vary the background with pairs that vary a co-occurring object raises worst-group accuracy above that of either pair type alone.

Machine learning models are inherently bound to the distribution of the training data, often exploiting non-causal shortcuts. As a result, achieving robustness to spurious correlations remains a challenge. While existing approaches rely on data manipulation or re-weighting strategies to achieve robustness, they typically require dense group labels, multiple training domains, or specialized pre-processing. We propose Invariance Pair Guidance (IPG), a method to mitigate reliance on spurious correlations using a sparse set of counterfactual pairs. Unlike other methods demanding extensive supervision, IPG utilizes a novel dual-update mechanism to dynamically correct the optimization trajectory. We generate input pairs that isolate the spurious attribute to define the invariance, a characteristic that should not affect the outcome of the model. Based on these pairs, we define a corrective gradient that complements the traditional gradient descent approach. The correction adapts via a predefined invariance condition. Experiments on ColoredMNIST, Waterbirds-100, and CelebA datasets demonstrate the effectiveness of our approach and its robustness to group shifts, supported by a theoretical convergence analysis. IPG offers a data-efficient and theoretically grounded path to robustness.